Flock Blames Bad Data While Blowing Past Every Guardrail

Two reporters have now been stopped for the same “stolen plate” alert on two different Land Rover press vehicles. The first was Joel Feder for The Drive. He “ended with four cop cars boxing [him] in,” after Plymouth PD tracked him “for days using Flock license plate cameras.” Tim Esterdahl, of Pickup Truck +SUV Talk, was stopped for 90 minutes while reviewing another Land Rover press vehicle shortly after.

The plate that sent the fuzz after Feder was a New Jersey manufacturer plate: 34 10 DTM. The plate that caused Esterdahl’s 90-minute stop was 34 08 DTM. On these plates, the small middle digits identify the individual car; the rest identifies the manufacturer’s fleet allocation.

Enter Joshua Thomas, Flock’s Chief Communications Officer, to explain the issue to The Drive:

Thomas says the plate was read correctly:

The way that the ML works is it correctly read what it was supposed to read… It was asked, can you find this? And it did find that.

But that police like to enter partial license plates:

The way that law enforcement likes to use these tools is, if any of the characters that they have put into these hot lists get read, they want to get those alerts.

According to Thomas, someone purposely told the system to alert on all results for a plate starting with “34” and ending in “DTM”.

That would be fine and dandy,^[1] were it not for NCIC policies, procedures, and technical constraints.

According to the reporting across the two incidents, a third plate from the same fleet — 34 03 DTM — was reported stolen in California (Feder later learned it had actually been lost by Land Rover during a photo shoot) and entered into NCIC as 34DTM, presumably into the License Plate File, where stolen plate entries go.^[2]

The Drive initially reported the entry came from LAPD; LAPD later said it was actually someone with the Los Angeles County Sheriff’s Department. LASD hasn’t commented.

The Drive says that “speaks to how messy and hard it is to ensure the data will always be clean going into a system like Flock—what it needs is stronger guardrails within it.” What The Drive likely doesn’t know is that those guardrails already exist: every single policy and procedure to prevent these issues from happening was violated.

First: the plate was lost, not stolen. The very first criterion for entry:

A uniquely-numbered stolen license plate can be entered only if a theft report has been made. … Agencies must have a theft report (electronic or hard copy) on file to support a license plate entry. Only the agency holding the theft report and having primary jurisdiction over the place of actual theft may make an NCIC entry

So either no qualifying theft report exists — making the record ineligible from the start — or a deputy took a theft report for a plate that nobody stole. Either way, that guardrail didn’t do much.

Second: the manual flatly prohibits what was entered. “Partial license plates must not be entered” (§ 2.5.2). The manual does handle two edge cases that resemble partial entries: plates longer than ten characters (enter the first ten, put the full number in the MIS miscellaneous field), and plates whose hyphens or symbols could cause a collision when stripped (“an example is plate number 1-1568 which would be matched with plate 11-568”). Again, the full plate goes in the MIS field. In both cases, the fix is more information in the record, never less.

And here’s the part that should embarrass Flock: NCIC’s one true partial-matching behavior comes with a warning label. When a plate exceeds eight characters, “the NCIC LIC search method uses the leftmost eight characters of the LIC” and the system automatically appends a caveat to the hit response: RECORD NIC/V123456789 IS BASED ON A PARTIAL LIC SEARCH - VERIFY ALL DATA BEFORE TAKING ACTION BASED ON THIS RESPONSE (§ 5.4, Vehicle File chapter).

When the FBI’s forty-year-old mainframe deviates from exact matching, it tells the officer, in capital letters, on the hit itself. Flock’s substring alerts arrive on someone’s smartphone without that information.

Third: the record must say what kind of plate it is. License Plate Type (LIT) is a mandatory entry field, and NCIC classifies everything that isn’t a standard passenger plate — including manufacturer plates — as a “special plate” with its own required type code (§ 1.1, § 2.5.5). A properly-built record would have announced itself as a New Jersey manufacturer plate.

Fourth: someone is supposed to look. The stopping agency must contact the entering agency, and:

“To confirm the hit” means to verify that the theft report is still outstanding; to verify that the license plate inquired upon is identical to the license plate described in the record…

Identical. Not “contains the same characters.” 34 10 DTM is not identical to 34 DTM, and no full plate ever can be identical to a truncated one. Applied as written, stops based on this record are impossible.

Fifth, the backstop: a “state Department of Motor Vehicles (DMV) where the plates are issued may compare license plate data with records contained in their files” and notify the agency or NCIC to correct or cancel the record (§ 1.4). It’s optional, and it requires an actual registry lookup rather than a format check — at five characters, “34DTM” is a facially valid NJ personalized-plate combination — but the mechanism exists. It would take milliseconds for NJ DMV to check if “34DTM” is an issued vanity plate, but the check was apparently not run.

We don’t really know exactly what the deputy entered. “34DTM” is the logical conclusion, because the license plate field “[m]ust not contain hyphens, spaces, or symbols.” Thomas’ suggestion that partial matching is a feature raises technical questions, because NCIC doesn’t have a “partial match” option. It prohibits partial plate entries, and the license plate field doesn’t allow wildcard entries like 34*DTM or 34%DTM.^[3] Any fuzzy matching happened in Flock’s system, against Flock’s copy of the hotlist.

We know 3408DTM, and 3410DTM triggered the alerts. Would 1234DTM or 34DTMABC have led to the same outcome? Would 314DTM have? How does Flock decide what’s “close enough” to send cops after a plate?^[4] And why isn’t it respecting NCIC’s constraints?

After his stop, Feder wrote, “the officers literally showed me the Flock app on their phone and explained how they used it.” Presumably, whatever the app displayed^[5] did not include the NCIC-required information that would have let them confirm the match, because the stop was improperly executed without confirmation — twice.

The guardrails and safeguards and backstops were all here. Entry requires a theft report. Partial plates are prohibited. The record must declare its plate type. Hits must be confirmed as identical with the entering agency before anyone acts. The DMV can flag and kill invalid records. Every layer of that system assumes exact matching, and several of those layers enforce it in writing.

Flock received an improperly-entered record for a plate that was never stolen, matched it against two different plates it read correctly — per Thomas — and called both a hit. The FBI’s rules say a hit means identical, but Flock says a hit means whatever it deems close enough.

Even with all those guardrails, four squad cars surrounded a journalist in a Kohl’s parking lot, and another reporter was detained for 90 minutes. Feder wrote that the cascade of errors “could’ve led to someone getting hurt or even killed.” Even Flock’s Thomas conceded the point, telling him: “I’m really thankful that your kids weren’t in the car. I’m thankful.”

We don’t need more guardrails. We need less Flock.